Hi Rob,
On 20/08/19 17:44, Rob Herring wrote:
>>>> + - i2c-alias-pool: list of I2C addresses that are known to be available on the
>>>> + "local" (SoC-to-deser) I2C bus; they will be picked at
>>>> + runtime and used as aliases to reach remove I2C chips
>>>
>>> s/remove/remote/
>>
>> Will fix.
>>
>>> Needs a vendor prefix.
>>
>> The ultimate goal here is to define a standard property that all chips
>> able to to I2C forwarding (video serdes or, potentially, other chips)
>> can use. That's why the GMSL (Maxim deser) developers are in Cc: they
>> are also facing a similar need.
>>
>> However I'm OK to change this to "ti,i2c-alias-pool" just in case there
>> are reasons to not use a common name [yet]. However, following this
>> argument, shouldn't a prefix be needed also for other nonstandard
>> strings, such as "i2c-atr" below?
>
> Okay, no vendor prefix is fine.
>
> 'i2c-atr' is a node name, not a property so no vendor prefix.
I see, thanks.
>>>> + - gpio-controller
>>>> + - #gpio-cells: must be 3: FPD-Link 3 RX port number, remote gpio number, flags
>>>
>>> We're pretty standardized on 2 cells for GPIO. Perhaps combine the port
>>> and gpio number to 1 cell.
>>
>> Oh dear. I dislike implementing software that does not model the
>> physical reality. I know it will bite me back sooner or later. Here we
>> _really_ have N physically separate GPIO controllers, and the number of
>> GPIOs they have depends on the model of the chip that is connected remotely.
>>
>> This is how things look in the case of 2 ports:
>>
>> <-- base board --> <------- remote camera module 1 ----->
>> .---------------------.
>> .-----. .------. | SER 1 |
>> | CPU |----|port 1|========|----------. |
>> `-----' | | FPD | GPIO ctl | |
>> | | Link 3 `---------------------'
>> | | cable ||||
>> | DES | remote GPIO pins
>> | |
>> | | <------- remote camera module 2 ----->
>> | | .---------------------.
>> |port 2| | SER 2 |
>> | |========|----------. |
>> `------' FPD | GPIO ctl | |
>> Link 3 `---------------------'
>> cable ||||||||
>> remote GPIO pins
>>
>> Perhaps we should have N separate gpiochips, one per port?
>
> Yes, seems like it.
I'm more and more convinced of that.
>>>> +Required subnodes:
>>>> +
>>>> + - ports: A ports node with one port child node per device input and output
>>>> + port, in accordance with the video interface bindings defined in
>>>> + Documentation/devicetree/bindings/media/video-interfaces.txt. The
>>>> + port nodes are numbered as follows:
>>>> +
>>>> + Port Description
>>>> + ------------------------------------
>>>> + 0 Input from FPD-Link 3 RX port 0
>>>> + 1 Input from FPD-Link 3 RX port 1
>>>> + 2 CSI-2 output
>>>> +
>>>> + Each port must have a "remote-chip" subnode that defines the remote
>>>> + chip (serializer) with at least a "compatible" property
>>>
>>> We don't allow other nodes within graph nodes. I'm not really clear what
>>> you are trying to do here.
>>
>> Each of the deser ports (2 ports in this chip) creates a physical
>> point-to-point "bus". It's called "FPD-Link 3" in the TI chips, "GMSL"
>> in the Maxim chips. One "remote chip" serializer can be connected to
>> each bus. The remote chip has, at least, a model (e.g. DS90UB953) and
>> some properties. So I need a place where model and properties can be
>> described. The port node looked like a good place, but as you point out
>> it is not.
>>
>> Adding to the above discussion about 3 gpio-cells, I can think of a
>> different, tentative DT layout:
>>
>> deser: deser@3d {
>> compatible = "ti,ds90ub954-q1";
>> reg-names = "main", "rxport0", "rxport1", "ser0", "ser1";
>> reg = <0x3d>, <0x40>, <0x41>, <0x44>, <0x45>;
>> clocks = <&clk_25M>;
>> interrupt-parent = <&gic>;
>> interrupts = <3 1 IRQ_TYPE_LEVEL_HIGH>;
>> reset-gpios = <&gpio_ctl 4 GPIO_ACTIVE_LOW>;
>>
>> i2c-alias-pool = /bits/ 16 <0x4a 0x4b 0x4c 0x4d 0x4e 0x4f>;
>>
>> ports {
>> #address-cells = <1>;
>> #size-cells = <0>;
>>
>> port@0 {
>> reg = <0>;
>> endpoint {
>> remote-endpoint = <&sensor_0_out>;
>> };
>> };
>>
>> port@1 {
>> reg = <1>;
>> endpoint {
>> remote-endpoint = <&sensor_1_out>;
>> };
>> };
>>
>> port@2 {
>> reg = <2>;
>> endpoint {
>> data-lanes = <1 2 3 4>;
>> /* Actually a REFCLK multiplier */
>> data-rate = <1600000000>;
>> remote-endpoint = <&csirx_0_in>;
>> };
>> };
>> };
>>
>> remote-chips {
>
> I don't have a better suggestion for the location of this.
>
>> #address-cells = <1>;
>> #size-cells = <0>;
>>
>> remote-chip@0 {
>> reg = <0>;
>> chip {
>> compatible = "ti,ds90ub953-q1";
>
> Seems like this should be in the parent? It's the ds90ub953
> implementing the I2C bus, GPIO controller, etc.?
You're right: it's the ds90ub953 implementing i2c, gpio etc. The
(undoubtedly weird) DT structure originates from the hotplugging issue
that I explained in the cover letter. The compromise I found to handle
it is to describe the remote I2C adapters and GPIO controllers as if
they were part of the "local" domain (the ds90ub954). This is why they
are represented in DT outside the ds90ub953 node: the "chip" node would
be added by an overlay, the i2c and gpio nodes are always there.
But this corresponds to how I implemented it in the code. Let's see if
at least DT can have a better layout:
deser: deser@3d {
compatible = "ti,ds90ub954-q1";
reg-names = "main", "rxport0", "rxport1", "ser0", "ser1";
reg = <0x3d>, <0x40>, <0x41>, <0x44>, <0x45>;
clocks = <&clk_25M>;
interrupt-parent = <&gic>;
interrupts = <3 1 IRQ_TYPE_LEVEL_HIGH>;
reset-gpios = <&gpio_ctl 4 GPIO_ACTIVE_LOW>;
i2c-alias-pool = /bits/ 16 <0x4a 0x4b 0x4c 0x4d 0x4e 0x4f>;
ports {
#address-cells = <1>;
#size-cells = <0>;
port@0 {
reg = <0>;
endpoint {
remote-endpoint = <&sensor_0_out>;
};
};
port@1 {
reg = <1>;
endpoint {
remote-endpoint = <&sensor_1_out>;
};
};
port@2 {
reg = <2>;
endpoint {
data-lanes = <1 2 3 4>;
/* Actually a REFCLK multiplier */
data-rate = <1600000000>;
remote-endpoint = <&csirx_0_in>;
};
};
};
remote-chips {
#address-cells = <1>;
#size-cells = <0>;
remote-chip@0 {
reg = <0>;
/* dynamic */
compatible = "ti,ds90ub953-q1";
gpio-functions = <...>;
remote_i2c0: i2c {
reg = <0>;
#address-cells = <1>;
#size-cells = <0>;
};
gpio-controller;
/* 2 cells for _each_ gpiochip: */
/* remote gpio num, flags */
#gpio-cells = <2>;
};
remote-chip@1 {
reg = <1>;
/* similar to remote chip 0 */
};
};
};
Now things _look_ correct: I just removed the "chip" subnode and hoisted
its properties up, where the i2c and gpio nodes are. However this would
not correspond a currently doable implementation. The gpio- and
i2c-related nodes under remote-chip@<N> would be always instantiated in
the base board DT, while the "compatible" property would be
added/removed by overlays on hotplug. This would require the deser chip
driver to probe the remote serializer chip not when the device node
appears but when the compatible property appears inside it. This would
look weird on the device driver side, even though I understand very well
that a DT needs to be future-proof much more than a driver.
I'll have a more detailed look into this.
Also, in case this were not clear, no hotplugging is actually
implemented currently, mostly due to the lack of a complete
implementation of runtime DT overlay insertion and removal. I think work
is in progress on that side, but it doesn't look like it is done yet.
However I'm trying to be ready for it when it will be available.
>> remote_i2c0: i2c@0 {
>
> What does '0' mean here? At a given level, you can only have 1 number
> space of addresses.
Copy-paste leftover, will fix.
>>>> + - #address-cells = <1>;
>>>> + - #size-cells = <0>;
>>>> +
>>>> + Optional properties for "i2c-atr" child bus nodes:
>>>> + - Other properties specific to the remote hardware
>>>
>>> Such as?
>>
>> "clock-frequency" at least. The remote chip is an I2C master, thus any
>> property that applies to an I2C master might apply as well.
>>
>> "clock-frequency" is only half-implemented in these RFC patches:
>>
>> * in patch 5, function ds90_rxport_add_serializer() passes the REFCLK
>> value of the deser to the remote chip (serializer): this value is
>> the physical reference clock the remote chip receives, all of its
>> timings are based on it
>>
>> * the remote chip, given refclk, computes the register values that best
>> approximate "clock-frequency" [not implemented in this version, would
>> be in patch 6]
>>
>> I plan to implement and document the whole clock-frequency feature for
>> the next patch iteration. But I'd love to receive comments about how
>> reflck is passed from the deser to the serializer via platform_data.
>
> Okay. Please try to make bindings as complete as possible even if
> there's not yet driver support.
>
>>>> + i2c-atr {
>>>> + #address-cells = <1>;
>>>> + #size-cells = <0>;
>>>> +
>>>> + remote_i2c0: i2c@0 {
>>>> + reg = <0>;
>>>> + #address-cells = <1>;
>>>> + #size-cells = <0>;
>>>
>>> Presumably, there are child I2C devices here. Please show that in the
>>> example.
>>
>> They are shown below to avoid excessive nesting, look for "&remote_i2c0"
>> (1).
>
> I'd rather have longish lines than have things split up.
Ok, but do we have 80-chars limit in the bindings too?
>>>> +&ds90ub954_fpd3_in0 {
>>>> + remote-chip {
>>>> + compatible = "ti,ds90ub953-q1";
>>>> + gpio-functions = <DS90_GPIO_FUNC_OUTPUT_REMOTE
>>>
>>> Not documented.
>>
>> That's because this node describes the remote chip. The remote chip is a
>> ti,ds90ub953-q1, its bindings are defined in patch 4. It's similar to
>> showing some child I2C devices under the I2C master, as you suggested
>> just above, except it's not an I2C bus but an FPD-Link 3 "bus".
>
> I think I realized this after sending this...
>
>>
>> Is it OK if I replace the whole gpio-functions node with the comment
>> "/* properties specific to the ti,ds90ub953-q1 chip */"?
>
> Not necessary. gpio-functions does need a vendor prefix though.
Fair enough. I think other models/brands of serdes chips will need
something similar, but they could accept different values, wo I'll add a
"ti," prefix.
--
Luca
